Adaptive weights learning in CNN feature fusion for crime scene investigation image classification

The combination of features from the convolutional layer and fully connected a neural network (CNN) provides an effective way to improve performance crime scene investigation (CSI) image classification. However, in existing work, as weights feature fusion do not change after training phase, it may produce inaccurate which affect classification results. To solve this problem, paper proposes adaptive method based on auto-encoder accuracy. includes following steps: Firstly, CNN model is trained by transfer learning. Next, convolution are extracted respectively. These then passed into for further learning with Softmax normalisation obtain performing final Experiments demonstrated that proposed achieves higher CSI compared fix fusion.